The Tencent Hunyuan team launched a quantitative version of the flagship model Hy3 (295B
parameters). The 1-bit version (IQ1_M) compresses the weight from 598 GB to 85.5 GiB, a 6.7-fold reduction, allowing deployment with a single 96GB inference graphics card; The 4-bit version (Q4_K_M) has a volume of 169.9 GiB, supporting two graphics cards. The quant version performs almost fully on tasks such as agents, multilingual code, tool calls, and long-form text understanding. The team also offers a GPTQ Int4 version, supporting vLLM deployment. With MTP speculative decoding, the 1-bit version gains decoding speed by about 50%, while the 4-bit version improves by nearly 60%. All versions are open-sourced and packaged in GGUF format, adapted for the llama.cpp ecosystem.
腾讯混元团队为旗舰模型 Hy3(295B 参数)推出量化版本。
1bit 版本(IQ1_M)将权重从 598 GB 压缩至 85.5 GiB,缩小 6.7 倍,单张 96GB 推理显卡即可部署;
4bit 版本(Q4_K_M)体积 169.9 GiB,两张显卡可承载。
量化版在 Agent、多语言代码、工具调用、长文理解等任务上表现接近满血模型。
团队还提供 GPTQ Int4 版本,支持 vLLM 部署。
配合 MTP 投机解码,1bit 版本解码速度提升约 50%,4bit 版本提升近 60%。
所有版本已开源并打包为 GGUF 格式,适配 llama.cpp 生态。
The readable text on this page was extracted from the public source and organized with attribution, publication time and the original link. Copyright remains with the original author and publisher.